FOREST FIRE DETECTION USING CONVOLUTIONAL NEURAL NETWORKS: A

DEEP LEARNING APPROACH

1Srikanth.Cheemaladinne, 2 Mr. Yerrabathana Guravaiah
1
M-Tech, Dept. of CSE Gokula Krishna College of Engineering, Sullurpet
2
Associate Professor, M-Tech., (Ph.D), CSE Gokula Krishna College of Engineering,
Sullurpet
Abstract: Forest fires are significant contributors to environmental degradation, causing
adverse impacts such as loss of wildlife habitat, extinction of plant and animal species, and
depletion of biodiversity and forest resources. To mitigate these effects, we propose a model
leveraging Convolutional Neural Networks (CNNs) for forest fire detection. CNNs are
effective for image classification tasks, capable of identifying key features within images to
detect specific objects. Our approach aims to identify the presence of a forest fire in an image
by training a CNN on a dataset labelled as "fire" and "no fire," where "fire" images depict
visible flames and "no fire" images show fire-free scenes. We utilize Keras and TensorFlow
libraries, which provide high-level APIs, simplifying the design and training of our model for
accurate and efficient fire detection.
Keywords: Forest fire detection, Convolutional Neural Network (CNN), Image classification,
Environmental monitoring.
1. Introduction human communities and animals. Fires
Wildfires are becoming more common have been identified using conventional
worldwide due to climate change, resulting image processing techniques that extract
in significant ecological harm and financial properties like colour, shape, and dynamic
losses [1]. The wildfire caused the texture. A fire recognition technique based
destruction of nearly 3,000 homes. The on logistic regression and temporal
toxic fumes from the wildfire, which killed smoothing was presented by Kong et al. [4]
450 people, injured over 4,000, and cost and utilised. Logistic regression was
over two billion dollars in medical care, had utilised to detect fires based on their size,
an impact on the world population. In motion, and colour features, while
addition to killing 3 billion wild animals, background removal was utilised to extract
the flames released around 7000 million colour component ratios as well as motion
metric tonnes of carbon dioxide and cost cues of flames to indicate probable fire
billions of dollars in damage [2]. locations. Mei et al. [5] presented a novel
Along with affecting human livelihoods, approach to fire identification that
regional economics, and environmental employed Hu characteristics and fire
health, fire also impacts forest ecosystems' overlap rate and motion severity rate across
environment, composition of species, and each frame to detect fires after initially
ecosystem design [3]. Therefore, it is extracting suspected fire zones using
crucial to watch wildfires intelligently. random forest, support vector machine, and
widespread harm is difficult for ecosystems enhanced ViBe algorithms. In order to
to recover from, which results in habitat and detect fire smoke, Dimitropoulos et al. [6]
biodiversity loss. Moreover, the debris and fused the features using an adaptive
ash have the potential to pollute water fractional union technique and then classed
supplies, which would further affect both them using SVM.

Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 164
Nevertheless, conventional image modest database of 1800 photographs was
processing techniques depend on used for this experiment.
preexisting knowledge to identify the fire Rahul et al. proposed a model that uses deep
feature extraction algorithms. Furthermore, learning methods to predict forest fires
the morphology of fires can reveal specific early. This study proposed a convolutional
characteristics at different stages, posing a neural network for image recognition to
challenge for a single fire feature extraction swiftly identify forest fires. This study
algorithm to adapt to diverse fire scenarios, employs a model that addresses this
thereby limiting its generalization ability. challenge through learning transfer. The
2. Literature Survey system utilizes and preprocesses the image
Tran et al. presented a forest fire response data. The system produces images using
model. This work focusses on augmentation methods like shearing.
implementing a wildfire armed guard with Models such as ResNet50 and VGC16,
damage estimate and detection. They trained to separate the data into two groups
implemented DetNAS uses neural fire and non-fire categorize the
architecture to find the optimum foundation photographs. They train the proposed
for object authentication. Approximately model using Resnet50 and transfer
400 thousand fire images are used for learning. By leveraging pre-trained
training and testing. A Bayesian neural networks, the system can achieve
network estimates damage. After detecting significant improvements in performance
forest fires, AI-enabled CCTV cameras with limited computational resources.
may provide real-time data to a central Fuzzy entropy optimised thresholding and
server. The server sends a UAV to check the STN-Based CNN are the tools that Avula et
fire site for damage. To seamlessly monitor al. suggest for the detection of forest fires.
and display the fire, the regression model For the purpose of smoke and forest fire
takes data from the burning region. UAV detection thresholding, this study suggests
and segmentation can assess damage area. a CNN-based model that is fuzzy entropy
The fundamental drawback of this response optimised. Great performance and accuracy
system is the challenge of real-time are achieved by integrating the CNN layer's
monitoring. spatial transformer network (STN) with the
For forest fire detection, Arteaga et al. SoftMax layer's entropy function adaptive
proposed deep learning model. This work threshold operation. As a critical step in the
evaluated the CNN models using pretrained pre-processing of video quality, lowering
forest fire pictures for economic growth false alarm rates is the primary objective of
devices like the Raspberry Pi. Forest fire this research.
prediction methods are offered. Two pre- 3. Proposed framework
trained CNN models can be utilised for this The proposed method makes use of a
assignment and evaluation. ResNet and two convolutional neural network's benefits.
pre-trained CNN families recommend 8 After receiving input, the CNN
models and 5 kinds. This research's main preprocesses it and pools the results using
limits are that the Resnet152 model may be the region of suggestions. Then, using
employed on a Raspberry Pi 3 Model B if convolutional layers, CNN's region-based
photos are not continually processed. A object identification algorithm divides
those predictions into fire and non-fire
Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 165
categories inside the ROI. This allows for a
more accurate and efficient detection of fire
occurrences within images. By leveraging
the hierarchical feature extraction
capabilities of CNNs, the method enhances
overall performance and reduces false
positives in identifying fire-related events.
Figure 1 mentions the CNN framework in
the proposed work. This section elaborates
the steps of the proposed framework. A) Fire images B) Non-fire images
Figure 2. Samples of the dataset
3.2 Image pre-processing
The next step in creating a high-quality
deep learning model is data preparation.
Here, we cleanse, process, or simply make
the data usable. As part of data preparation,
we eliminate noise and other unwanted
items from the picture. The algorithm needs
pertinent data to function properly;
otherwise, it can provide undesirable
Figure 1. Proposed CNN framework outcomes. To ensure the model's accuracy
3.1 Dataset collection and reliability, it is crucial to also perform
The data set that was made available by data augmentation, which helps to
Brsdincer (2021). The purpose of this data artificially expand the dataset by creating
collection was to train a model that could variations of the existing data. This process
differentiate between photographs that not only enhances the model's ability to
contain fire (fire images) and images that do generalize but also mitigates the risk of
not contain fire (non-fire images). A total of overfitting the training data.
999 photographs are separated into two 3.3 Model Architecture
directories, with the “fire images” folder Convolutional Layers: To assist the
including 755 images of fires that occur model, differentiate between "fire" and "no
outside as well. There are some of them that fire" images, use many convolutional layers
have strong smoke, while the other one is to learn characteristics from the images,
called “non fire images”, and it has 244 such as forms, edges, and textures.
images of nature (such as forests, trees, Pooling Layers: To minimise spatial
grass, rivers, and so on). The sample of the dimensions, minimise computation, and
dataset in mentioned in Figure 2. preserve important characteristics, add
pooling layers (like MaxPooling) after
convolutional layers.
Completely Interconnected Layers:
Convolutional layer output should be
flattened before being linked to fully
connected (dense) layers, which execute the

Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 166
final classification by combining deep learning model. In the validation
information learnt in earlier layers. phase, a separate set of images, distinct
Output layer: According to the variety of from those used during model training, is
classes, use either a SoftMax or sigmoid used to evaluate the model’s performance.
activation function in the output layer This process helps to confirm the model’s
(SoftMax for multiple classes, sigmoid for ability to generalize beyond the training
binary classification). dataset and provides insights into its
3.4 CNN Model reliability in detecting forest fires under
To create distinct compact feature maps, the diverse conditions.
recommended CNN model employs 4. Results
distinct filters in the convolution and The initial series of studies was conducted
pooling layers. To assess the fire's using the Forest Fire database, which
confidence percentage, we employ the comprises 2 participants and 999 photos in
Squeeze Net Model. The precision and total. One image from each participant was
minimal processing power of the Squeeze designated as the training set, while the
Net are well-known. Compared to other remaining images constituted the testing set
training models, it produces more accurate for each iteration. The experiment was
outcomes. In order to give the convolution conducted ten times to utilise each image as
layer bigger activation maps, it down the training set for assessment. In the
samples the input channels. The foundation evaluation, the CNN algorithm got 93% of
of the Squeeze Net paradigm consists of fire accuracy for 10 iterations. The experiment
modules. It consists of two pinching and results of the CNN are mentioned in Figure
expanding layers. A couple pooling layers 3.
and a stack of fire modules make up a
squeeze net. The size of the feature maps in
the enlarged and squeezed layers is same.
While the extended layer increases depth,
the squeezing layer diminishes it.
Significant characteristics are
automatically learnt by deep CNN models
from datasets. Eliminating preprocessing
and finding patterns—both necessary for
fire danger detection—are further
motivators. Two classes—fire and
Figure 3. A) Figure 3. B)
normal—are used to train the frames. There
Accuracy Accuracy loss
is no need for additional steps if the frames
validation
are non-fire. Binarization comes after
Figure 3. Evaluation results of CNN
additional feature extraction if the frame is
Sample images demonstrate that the model
determined to be fire. The probability
effectively distinguishes between "fire" and
ratings of the network are used to assign the
"no fire" scenarios, indicating its ability to
classes.
identify essential fire-related
3.5 Validation and testing
characteristics, such as smoke and flames,
Validation and testing are essential steps to
across diverse contexts. Figure 4.
assess the accuracy and functionality of the
Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 167
demonstrated the prediction results of the variety of datasets with different fire
CNN algorithm. circumstances.
References
[1] Xu, R.J.; Lin, H.F.; Lu, K.J.; Cao,
L.; Liu, Y.F. A Forest Fire Detection
System Based on Ensemble
Learning. Forests 2021, 12, 217.
[2] Celermajer, D.; Lyster, R.; Wardle,
G.M.; Walmsley, R.; Couzens, E.
The Australian bushfire disaster:
How to avoid repeating this
Figure 4. Prediction Results catastrophe for biodiversity. Wiley
This experiment was conducted using the Interdiscip. Rev.-Clim. Chang.
Python programming language with the 2021, 12, e704.
Flask framework on a Windows 10 [3] Ma, W.Y.; Feng, Z.K.; Cheng, Z.X.;
machine, equipped with an Intel(R) Core Chen, S.L.; Wang, F.G. Identifying
(TM) i5 processor at 2.3 GHz and 16 GB of Forest Fire Driving Factors and
RAM, utilising hardware acceleration for Related Impacts in China Using
model training. Random Forest Algorithm. Forests
5. Conclusions 2020, 11, 507.
Using image data, we created a [4] Kong, S.G.; Jin, D.; Li, S.; Kim, H.
Convolutional Neural Network (CNN) Fast fire flame detection in
model in this study to identify forest fires. surveillance video using logistic
CNNs are effective in differentiating regression and temporal smoothing.
between "fire" and "no fire" situations, as Fire Saf. J. 2016, 79, 37–43.
seen by our model's high accuracy of 93% [5] Mei, J.J.; Zhang, W. Early Fire
on the test dataset. We made it easier to Detection Algorithm Based on ViBe
create and train our model by utilising the and Machine Learning. Acta Opt.
Keras and TensorFlow frameworks, which Sin. 2018, 38, 0710001.
allowed for precise and rapid [6] Dimitropoulos, K.; Barmpoutis, P.;
categorisation. Grammalidis, N. Higher Order
The model is a useful tool for early forest Linear Dynamical Systems for
fire detection and monitoring, since the Smoke Detection in Video
findings show that it can accurately detect Surveillance Applications. IEEE
outward indicators of fire. In order to lessen Trans. Circuits Syst. Video Technol.
the negative effects of forest fires on the 2017, 27, 1143–1154.
environment, this model may help in real- [7] D. Q. Tran, M. Park, Y. Jeon, J. Bak
time applications by enabling quick and S. Park, "Forest-Fire Response
response times and offering early warnings. System Using Deep-Learning-
Future developments may include Based Approaches with CCTV
strengthening robustness, including more Images and Weather Data," in IEEE
environmental data for a more thorough fire Access, vol. 10, pp. 66061-
risk assessment, and training on a wider 66071,2022,doi:10.1109/ACCESS.
2022.3184707.
Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 168
 [8] Arteaga, M. Diaz and M. Jojoa, [10] S. B. Avula, S. J. Badri and
"Deep Learning Applied to Forest G. Reddy P, "A Novel Forest Fire
Fire Detection," 2020 IEEE Detection System Using Fuzzy
International Symposium on Signal Entropy Optimized Thresholding
Processing and Information and STN-based CNN," 2020
Technology (ISSPIT), 2020, pp. 1- International Conference on
6, doi: Communication Systems &
10.1109/ISSPIT51521.2020.94088 Networks (COMSNETS), 2020, pp.
59. 750-755, doi:
[9] M. Rahul, K. Shiva Saketh, A. 10.1109/COMSNETS48256.2020.9
Sanjeet and N. Srinivas Naik, 027347.
"Early Detection of Forest Fire [11] Brsdincer. (2021). Forest
using Deep Learning," 2020 IEEE Fire Detection-Prediction/ALL
REGION 10 CONFERENCE PROCESS. Kaggle.
(TENCON), 2020, pp. 1136-1140, https://www.kaggle.com/code/brsdi
doi: ncer/forest-fire-detection-
10.1109/TENCON50793.2020.929 prediction-all-process/input
3722.

Volume 14, Issue 11, Nov 2024 ISSN 2457-0362 Page 169